On April 17, 2026, the Federal Reserve, the OCC, and the FDIC issued the first overhaul of model risk management guidance in fifteen years. SR 26-2 replaces SR 11-7, the framework that has shaped how banks build, validate, and govern models since 2011.
Buried in a footnote is the line that matters most to anyone running generative AI or AI agents inside a bank: those systems are explicitly outside the scope of the new guidance.
That should not be read as a green light. It’s the opposite. Examiners are already asking every bank, regardless of size, how they govern the AI systems the rule doesn’t cover. And a growing body of evidence suggests most institutions aren’t ready for that conversation. This post breaks down exactly what SR 26-2 says, what’s happened since it was issued, and what banks need to put in place now, before the next exam and before the regulatory guidance written specifically for GenAI and agentic AI arrives.
What SR 26-2 Actually Says
SR 26-2 preserves the core discipline that has governed bank models for over a decade: validated governance, independent review, and documented assumptions for the deterministic and near-deterministic systems used in credit scoring, market risk, and regulatory capital calculations. If your bank runs a traditional quantitative model, the expectations you’re used to under SR 11-7 largely carry forward.
The scope footnote is where the guidance draws its line: it states that generative AI and agentic AI models are novel and rapidly evolving, and as a result are not within the scope of the guidance. It goes on to instruct banks to rely on their own risk management and governance practices to determine appropriate controls for any tools or systems the document doesn’t cover, while confirming that the guidance’s principles do apply to traditional statistical models and to non-generative, non-agentic AI.
Two things follow from that language. First, the regulators aren’t saying GenAI and agentic AI are unregulated: they’re saying the specific technical requirements in SR 26-2 weren’t built for them, and banks are on the hook for filling that gap themselves. Second, the agencies have been explicit in the SR 26-2 letter itself that SR 26-2 creates no new enforceable requirements: failing to meet it will not, on its own, trigger supervisory criticism, though supervisory action can still follow from violations of law or unsafe practices. There is no checklist yet. There is only the expectation that banks build one.
What’s Happened Since April
The gap the footnote created hasn’t stayed theoretical. Two developments since SR 26-2 was issued show how quickly it’s becoming a live exam issue.
Every routine exam now includes AI governance questions. Reporting on the shift shows that examiners at the Fed, OCC, and FDIC have started embedding AI oversight questions into standard bank audits, not as a special AI review, but as part of business as usual. The questions cluster around three failure modes: whether AI systems can access or infer data outside their authorized scope, whether the bank can actually shut down a system that starts behaving unexpectedly, and whether third-party AI vendors, and their subcontractors, are held to the same governance standard the bank itself is expected to meet. That last point matters more than it sounds: a vendor’s AI tool may itself sit on top of a foundation model provider, and most third-party risk frameworks weren’t built to see that deep into the chain.
Most banks can’t answer those questions with confidence. A Wolters Kluwer survey of 230 banking professionals across community, midsize, and large institutions, published June 10, 2026, found that 72% identified kill-switch protocols or regulatory reporting of AI failures as their least-prepared area. Put plainly: nearly three in four banks aren’t confident they could shut down a malfunctioning AI model or tell a regulator what went wrong if it happened tomorrow. As American Banker’s coverage of the survey noted, the survey’s authors called this a minimum viable requirement, not an advanced capability, and noted that banks are scaling AI deployment considerably faster than they’re building the governance and incident response architecture to support it.
The risk isn’t hypothetical. In July 2025, the Massachusetts Attorney General reached a $2.5 million settlement with Earnest Operations after its AI underwriting model was found to produce racially discriminatory loan terms. As ABA Banking Journal’s case summary details, the model used a college-level default-rate variable that functioned as an illegal proxy under the Equal Credit Opportunity Act, and no human reviewer caught it before real harm occurred. ECOA requires lenders to give specific, accurate reasons for a credit denial. A model that can’t explain its own decision can’t produce that legally required notice, and regulators have made clear that gap creates liability, not an exemption.
Why the Old Governance Playbook Doesn’t Transfer
The exclusion in SR 26-2 isn’t arbitrary. It reflects a real technical difference between the models the guidance was written for and the systems banks are deploying now.
A credit scorecard or fraud-detection model is deterministic, or close to it: fixed inputs produce traceable, reproducible outputs. Regulators can validate that kind of system against documented assumptions and out-of-sample data, which is exactly what SR 11-7-style governance was designed to do.
Large language models and agentic systems don’t behave that way. Their outputs are probabilistic: the same prompt can produce different responses depending on context, phrasing, or configuration. Retrieval-Augmented Generation, the architecture behind most enterprise GenAI assistants, compounds the problem. If access controls aren’t explicitly enforced at the retrieval layer, a query about one customer can inadvertently surface another customer’s data. Agentic AI raises the stakes further. Instead of just generating text, agents take autonomous multi-step action across systems (filing disputes, processing payments, updating records), often without a human checkpoint at every step, meaning a single authorization error can cascade through several consequential decisions before anyone notices.
None of this means GenAI and agentic AI can’t be governed. It means they need to be governed differently, with monitoring, documentation, and human oversight designed for probabilistic, autonomous behavior, not adapted from a checklist built for regression models.
The Interim Standard: Build the Framework Now
The agencies have signaled a separate request for information on GenAI and agentic AI is coming. There’s no published timeline. Waiting for it isn’t a viable strategy, not with examiners already asking the questions and 72% of banks unable to answer them.
The more defensible position is to apply the principles underneath SR 26-2 (inventory, monitoring, independent review, documentation, human accountability) to GenAI and agentic AI now, adapted to how these systems actually fail. In practice, that means a governance framework built around five pillars:
1. A complete AI and agent inventory. Every GenAI tool, LLM, and agent in production needs a documented owner, purpose, scope, and review cadence: the same starting point regulators expect for traditional models, extended to systems SR 26-2 doesn’t name.
2. Monitoring built for how GenAI and agents actually fail. Generic model monitoring (drift, accuracy) doesn’t catch what breaks in a language model or an agent. Banks need: – Adversarial prompting and prompt-injection testing, to find out how the system behaves when someone tries to manipulate it before a bad actor finds out first – Hallucination monitoring, especially in RAG pipelines where a model can generate a confident, fabricated answer grounded in nothing – LLM-as-a-judge evaluation, to score model outputs at scale in a way manual spot-checks can’t match – Data drift, fairness, and interpretability testing for the numeric components still embedded in most hybrid AI systems
3. Full source traceability. Every GenAI output should be traceable back to the source data or documents that produced it. This is the direct technical answer to the “data boundary enforcement” concern examiners are raising, and it’s also what makes an explainable adverse-action notice possible: the exact gap that created liability in the Earnest case.
4. Human-in-the-loop as a control, not a feature. This is where kill-switch capability actually lives. Human checkpoints need to sit at every material decision point in an agentic workflow, with clear accountability for who reviews what, and a documented path to halt or roll back a system that starts behaving unexpectedly, before a regulator asks whether one exists
5. Change control and audit-ready documentation. An indelible record of what changed, when, and why, produced as a byproduct of how the system runs, not reconstructed from chat logs and tickets after an examiner asks for it.
These five pillars are the standard to build toward. Putting them into practice runs into two specific risk surfaces most banks underestimate: the spreadsheets and end-user-computing files that increasingly feed AI systems, and the fundamentally different way GenAI and agents need to be validated compared to traditional models.
Want to see what closing this gap looks like in practice? Talk to CIMCON directly.
Key Takeaways
- SR 26-2 modernizes model risk governance for traditional quantitative models but explicitly excludes GenAI and agentic AI from its scope.
- The exclusion doesn’t mean banks are off the hook: regulators expect banks to apply their own governance practices, and examiners are already asking detailed AI governance questions in every routine exam.
- 72% of banks surveyed can’t confidently confirm they could shut down a malfunctioning AI model or report a failure to regulators.
- Real enforcement precedent already exists: a $2.5 million ECOA settlement over an AI underwriting model shows the legal exposure is current, not theoretical.
- The interim standard is to apply SR 26-2’s underlying principles (inventory, GenAI/agent-specific monitoring, source traceability, human-in-the-loop control, and change documentation) to systems the guidance doesn’t yet name.
FAQ
Does SR 26-2 apply to our chatbots and AI agents? No. SR 26-2 explicitly excludes generative AI and agentic AI from its scope. It applies to traditional quantitative models and non-generative, non-agentic AI. But the guidance instructs banks to apply their own governance practices to fill that gap, and examiners are already asking about GenAI and agent oversight in routine exams.
If GenAI and agentic AI are out of scope, are we actually exposed to regulatory risk right now? Yes. SR 26-2 creates no enforceable requirement on its own, but examiners are asking about kill-switch capability, data access boundaries, and vendor oversight for these systems in every exam. Existing consumer protection and fair lending laws, like ECOA, already apply regardless of whether a rule specifically names GenAI, as the Earnest Operations settlement shows.
What is an AI kill switch, and why do most banks lack one? It’s the ability to identify, stop, and report a malfunctioning AI system: documented rollback procedures, routing controls to divert traffic away from the model, and the ability to explain the failure before a regulator asks. Because AI models are probabilistic and often run on vendor-managed infrastructure, this is harder to build than a traditional software failover, and a 2026 survey found nearly three in four banks are unprepared in this area.
What’s the difference between validating a numeric model and validating an LLM or AI agent? Numeric models are validated against fixed assumptions and out-of-sample data because their behavior is deterministic or close to it. LLMs and agents are probabilistic and can produce different outputs from the same input, so validation has to include hallucination monitoring, adversarial prompt testing, source attribution, and, for agents, human checkpoints at each autonomous action.
What should a bank’s GenAI and agentic AI governance framework include today? At minimum: a complete inventory of every GenAI tool and agent in production, monitoring designed for how these systems actually fail (hallucination, adversarial prompting, LLM-as-a-judge evaluation), full source traceability, human-in-the-loop control at every material decision point, and audit-ready change documentation.
References
1.Board of Governors of the Federal Reserve System. “SR 26-2: Revised Guidance on Model Risk Management.” April 17, 2026. federalreserve.gov
2.Wolters Kluwer. “Wolters Kluwer Survey Highlights AI Governance and Other Ongoing Needs for Banking Institutions.” June 10, 2026. wolterskluwer.com
3.Crosman, Penny. “72% of Banks Lack AI Model ‘Kill Switches,’ Failure Reporting.” American Banker, June 10, 2026. americanbanker.com
4.Massachusetts Attorney General’s Office. “AG Campbell Announces $2.5 Million Settlement With Student Loan Lender For Unlawful Practices Through AI Use, Other Consumer Protection Violations.” July 10, 2025. mass.gov
5.ABA Banking Journal. “Mass. AG Reaches Settlement With Student Loan Firm for $2.5M Over AI Lending Bias.” 2025. bankingjournal.aba.com
